This project is a secondary data analysis of the Millennium Cohort Study [18]. Since 2001, eight data sweeps have been collected. At the first data collection point, 18,818 cohort children and 18,552 families were included. Further information about the MCS study characteristics and design can be found in Plewis and colleagues [19]. A study protocol for the current project was pre-registered on the Open Science Framework (OSF, https://doi.org/10.17605/OSF.IO/S92MT) prior to data analysis and all analysis codes uploaded on GitHub (https://github.com/M-COpitz/MCS_Sleep_Disordered_Eating.git). This study received ethical approval from the University of Edinburgh Research Ethics Committee (Reference: 22-23CLPS066). Further information on available data and measures is provided in the supplementary material.

MCS data from sweep 6 (N = 15,414 children aged 14, assessed 2015) and sweep 7 (N = 14,438 children aged 17, assessed 2018) were used for data analysis and accessed through the UK Data Service. All cohort children who self-reported on restrictive eating behaviours and sleep outcomes at both time points were included in the analyses. For twin pairs, one twin was randomly selected for inclusion. This led to a final analytic sample of N = 6,041. At age 14, a self-selected subsample of participants additionally reported on their time use on two randomly assigned assessment points. Those reporting on their time-use without missing data on both days were included for all sub-sample analyses (N = 2,164).

Restrictive eating behaviours were assessed using three available indicators: dietary restriction (“Have you ever eaten less food, fewer calories, or foods low in fat to lose weight or to avoid gaining weight?”, 1 = yes, 0 = no), exercise for weight control (“Have you ever exercised to lose weight or to avoid gaining weight?”, 1 = yes, 0 = no), and intention to lose weight (“Which of the following are you trying to do about your weight?”, 1 = lose weight, 0 = gain weight, stay the same weight, I am not trying to do anything about my weight). These items were assessed at age 14 and 17, with the former two items referring to life-time behaviours up until the age of 14 at the first assessment point, and recent experiences (previous 12-months) at age 17.

At age 14, the MCS assessed adolescent-reported bed-/wake time categories, and single-items on sleep onset timings and difficulties with awakening during the night. To depict problems with sleep and to replicate previous approaches, the following sleep variables were created: participants’ typical sleep durations on school days and school-free days were categorised as ≤ 8 h, 8–9 h, 9–10 h, and ≥ 10 h, to align categories with recommendations of the American Academy of Sleep Medicine for adolescent populations (13–18 years) [17]. The median wake-up and bedtimes were used to divide participants in two groups. As sleep variables were already pre-categorised and this study aimed to assess associations with problematic sleep behaviours, dichotomous variables were created to depict earlier/later (i.e., potentially problematic) wake-up (< 7.30am school days, < 10.30am school-free days) and bedtime categories (> 10.30pm school days, > 11.30pm school-free days). Social jetlag, the discrepancy between individuals’ biological and social clocks [20], was calculated as:

Social Jetlag = midpoint of sleep on school-free days (biological time) – midpoint of sleep on school days (social time).

In line with clinical recommendations [21] and previous approaches [22], sleep onset latency was defined as presence (> 30 min) or absence (≤ 30 min) of sleep onset difficulties. Wake After Sleep Onset (WASO) was equally dichotomised (presence = difficulties “all of the time”, “most of the time”, or “a good bit of the time”) [22]. Sleep problems were thereby contrasted with a lack of sleep problems in accordance with previous studies [22], to investigate links between restrictive eating and problematic sleep behaviours.

Time use data (TUD) captured a subsample’s activities during two 24 h periods (during the week and during the weekend), recorded within 10 days of the interview visit. Three variables were created for both weekdays and weekend days: 24-hour sleep duration (all sleep reported during each 24 h period), awakening during the night (the presence or absence of reporting at least one incidence of awakening between 8pm and 5.30am to avoid capturing naps and typical wake-up times), and day-time naps (presence or absence of naps during the day, with a nap being defined as any sleeping period between 9am and 7pm, which occurred after being awake for at least 60 min).

Participants’ self-rated overall sleep quality (“During the past month, how would you rate your sleep quality overall? Would you say it has been…”, “Very good”, “Fairly good”, “Fairly bad”, “Very bad”) was assessed at age 17.

Symptoms of Depression were assessed via the Short Moods and Feelings Questionnaire (SMFQ), a 13-item measure with sum-score values ranging from 0 to 23 [23]. Adolescents (age 14) self-reported how they were feeling/acting during the previous two weeks (e.g., “I felt miserable or unhappy”; “not true”, “sometimes”, “true”). Good internal reliability of the SMFQ was previously found across time within a general adolescent population [24]. Within the present study, Cronbach’s alpha was 0.93 (ω = 0.93). For all relevant analyses, depressive symptoms were modelled as one latent factor.

Relevant sociodemographic covariates were included as provided in the MCS. These were participants’ sex (female or male), ethnicity (White, mixed, Indian, Pakistani or Bangladeshi, Black or Black British, other ethnic group), weight category (adjusted BMI categories formulated by the UK90 [25]), and household income (equivalised income quintiles by country). Seasonality was accounted for in all TUD analyses (spring, summer, autumn, and winter).

All analyses were conducted using R version 4.3.1 (see supplementary material for packages used). Patterns of item missingness for predictors and outcomes were explored using Little’s MAR test and the ‘naniar’ package. At data sweep 7, 55.2% of the full initial MCS sample (N = 19,243) were considered ‘productive’ survey responders. For more information on participant response rates, see Ploubidis and Mostafa [26] for sweep 6 and Ipsos MORI [27] for sweep 7. Missing data and survey weighting approaches for this project are outlined in the supplementary material.

Associations between relevant variables were calculated using chi-square tests, phi coefficients (dichotomous variables), Cramer’s V (categorical-dichotomous associations), and point-biserial correlation coefficients (continuous-categorical associations). To assess the latent factor structure as well as the extent to which the specified latent construct measured restrictive eating behaviours consistently across time, the model’s longitudinal measurement invariance was evaluated following the approach outlined in Mackinnon et al. [28] (findings presented in supplementary material).

To address the outlined research questions, a series of regression analyses were conducted within a structural equation modelling (SEM) framework. Restrictive eating behaviours were operationalized as a latent factor, and all sleep variables were added as manifest variables. First, sleep indicators (measured at age 14) were individually modelled as predictors of restrictive eating behaviours (measured at age 17), controlling for restrictive eating behaviours at age 14. Second, restrictive eating behaviours at age 14 was specified as a predictor for perceived sleep quality (measured at age 17), controlling for all sleep indicators at age 14. All analyses that included TUD were conducted on the available sub-sample. Due to the non-continuous nature of the indicator variables, the robust weighted least squares (WLSMV) estimator was specified to estimate robust standard errors. SEM mediation models were tested for all significant longitudinal models, specifying the mediator as measured at baseline (age 14). A value of p <.05 was used to determine significance. To account for multiple testing, the Benjamini-Hochberg correction with a false discovery rate of 0.05 was used to adjust p-values in relevant analyses.